Tabarrok Study

THE FUGITIVE: EVIDENCE ON PUBLIC VERSUS
PRIVATE LAW ENFORCEMENT FROM

BAIL JUMPING*
ERIC HELLAND and ALEXANDER TABARROK
Claremont-McKenna College George Mason University
Abstract
On the day of their trial, a substantial number of felony defendants fail to appear.
Public police have the primary responsibility for pursuing and rearresting defendants
who were released on their own recognizance or on cash or government bail. De-
fendants who made bail by borrowing from a bond dealer, however, must worry
about an entirely different pursuer. When a defendant who has borrowed money skips
trial, the bond dealer forfeits the bond unless the fugitive is soon returned. As a
result, bond dealers have an incentive to monitor their charges and ensure that they
do not skip. When a defendant does skip, bond dealers hire bounty hunters to return
the defendants to custody. We compare the effectiveness of these two different systems
by examining failure-to-appear rates, fugitive rates, and capture rates of felony de-
fendants who fall under the various systems. We apply propensity score and matching
techniques.
I. Introduction
A pproximately one-quarter of all released felony defendants fail to appear

at trial. Some of these failures to appear are due to sickness or forgetfulness
and are quickly corrected, but many represent planned abscondments. After
1 year, some 30 percent of the felony defendants who initially fail to appear
remain fugitives from the law. In absolute numbers, some 200,000 felony
defendants fail to appear every year, and of these, approximately 60,000 will
remain fugitives for at least 1 year.1
* The authors’ names are in alphabetical order. We wish to thank Jonathan Guryan, Steve
Levitt, Lance Lochner, Bruce Meyer, Jeff Milyo, Christopher Taber, Sam Peltzman, and seminar
participants at Claremont McKenna College, the American Economic Association annual meet-
ings (2002), George Mason University, Northwestern University, and the University of Chicago.
1
These figures are from the State Court Processing Statistics (SCPS) program of the Bureau
of Justice Statistics and can be found in U.S. Department of Justice, Bureau of Justice Statistics,
Felony Defendants in Large Urban Counties (various years). We describe the data at greater
length below. The SCPS program creates a sample representative of 1 month of cases from
the 75 most populous counties (which account for about half of all reported crimes). In 1996,
the sample represented 55,000 cases, which in turn represent some 660,000 filings in a year
and 1,320,000 filings in the nation. The absolute figures are calculated using this total, and
[Journal of Law and Economics, vol. XLVII (April 2004)]

䉷 2004 by The University of Chicago. All rights reserved. 0022-2186/2004/4701-0005$01.50
93
94 the journal of law and economics
Defendants who fail to appear impose significant costs on others. Direct

costs include the costs of rearranging and rescheduling court dates, the wasted
time of judges, lawyers, and other court personnel, and the costs necessary
to find and apprehend or rearrest fugitives. Other costs include the additional
crimes that are committed by fugitives. In 1996, for example, 16 percent of
released defendants were rearrested before their initial case came to trial.2
We can be sure that the percentage of felony defendants who commit ad-
ditional crimes is considerably higher than their rearrest rate. We might also
expect that the felony defendants who fail to appear are the ones most likely
to commit additional crimes. Indirect costs include the increased crime that
results when high failure-to-appear (FTA) and fugitive rates reduce expected
punishments.3
The dominant forms of release are by surety bond, that is, release on bail
that is lent to the accused by a bond dealer, and nonfinancial release. Just
over one-quarter of all released defendants are released on surety bond, and
a very small percentage pay cash bail or put up their own property with the
court (less than 5 percent combined); most of the rest are released on their
own recognizance or on some form of public bail (called deposit bond) in
which the defendant posts a small fraction, typically 10 percent or less, of
the bail amount with the court.
Estimating the effectiveness of the pretrial release system in the United
States can be characterized as a problem of treatment evaluation. Treatment
evaluation problems can be difficult because treatment is rarely assigned
randomly. Release assignment, for example, is based on a judge’s assessment
of the likelihood that a defendant will appear in court as well as on consid-
erations of public safety. Correctly measuring treatment effects requires that
we control for treatment assignment. In this paper, we control for selection
by matching on the propensity score.4
We estimate the treatment effect for three outcomes—the probability that
the release, failure-to-appear (FTA), and fugitive (defined as FTA for 1 year or more) rates
from the random sample. See note 2 infra.
2
U.S. Department of Justice, Bureau of Justice Statistics, Felony Defendants in Large Urban
Counties, 1996 (1999) (http://www.ojp.usdoj.gov/bjs/pub/pdf/fdluc96.pdf).
3
Justice delayed can mean justice denied in practice as well as in theory. Thousands of
cases are dismissed on constitutional grounds every year because police fail to serve warrants
in a timely manner. See Kenneth Howe & Erin Hallissy, When Justice Goes Unserved:
Thousands Wanted on Outstanding Warrants—but Law Enforcement Largely Ignores Them,
S.F. Chron., June 22, 1999, at A1.
4
For the matching method, see Donald B. Rubin, Estimating Causal Effects of Treatments
in Randomized and Nonrandomized Studies, 66 J. Educ. Psychol. 688, 701 (1974); Donald B.
Rubin, Assignment to Treatment Group on the Basis of a Covariate, 2 J. Educ. Stat. 1, 26
(1977); Paul R. Rosenbaum & Donald B. Rubin, Reducing Bias in Observation Studies Using
Subclassification on the Propensity Score, 79 J. Am. Stat. Assoc. 516, 524 (1984); Rajeev H.
Dehejia & Sadek Wahba, Causal Effects in Non-experimental Studies: Re-evaluating the Eval-
uation of Training Programs (Working Paper No. 6586, Nat’l Bur. Econ. Res. 1998); James J.
Heckman, Hidehiko Ichimura, & Petra Todd, Matching As an Econometric Evaluation Esti-
mator, 65 Rev. Econ. Stud. 261, 294 (1998).
bail jumping 95
a defendant fails to appear at least once, the probability that a defendant

remains at large for 1 year or more conditional on having failed to appear
(what we call the fugitive rate), and the probability that a defendant who
failed to appear is recaptured as a function of time.
The earlier economic studies of the bail system examine the role of the
bail amount in the decision to fail to appear, generally finding that higher
bail reduces FTA rates.5 These studies did not focus on the central issue of
this paper—the different incentive effects of the various release types.6
II. History of Pretrial Release and Incentive Effects

of Release Systems
Although money bail is still the most common form of release, money
bail and especially the commercial surety industry have come under increas-
ing and often virulent attack since the 1960s.7 Bail began as a progressive
measure to help defendants get out of jail when the default option was that
all defendants would be held until trial. In the twentieth century, however,
the default option was more often thought of as release, and thus money bail
was reconceived as a factor that kept people in jail. In addition, the greater
burden of money bail on the poor elicited growing concern.8 As a result,
5
William M. Landes, The Bail System: An Economic Approach, 2 J. Legal Stud. 79, 105
(1973); William M. Landes, Legality and Reality: Some Evidence on Criminal Procedure, 3
J. Legal Stud. 287 (1974); Stevens H. Clark, Jean L. Freeman, & Gary G. Koch, Bail Risk:
A Multivariate Analysis, 5 J. Legal Stud. 341, 385 (1976); Samuel L. Myers, Jr., The Economics
of Bail Jumping, 10 J. Legal Stud. 381, 396 (1981).
6
Ian Ayres & Joel Waldfogel, A Market Test for Discrimination in Bail Setting, 46 Stan.
L. Rev. 987, 1047 (1994), demonstrates the subtlety of the distinctions made by bond dealers
in setting bail bond rates. Although the courts (in New Haven, Connecticut, in 1990) set higher
bail amounts for minority defendants than for whites, Ayres and Waldfogel find that bond
dealers acted in precisely the opposite manner. What this pattern suggests is that judges set
higher bail for minority defendants than for white defendants with the same probability of
flight. Bond dealers are then induced by competition to charge minorities relatively lower bail
bond rates.
7
Floyd Feeney, Foreword, in Bail Reform in America, at ix (Wayne H. Thomas, Jr., ed.
1976), for example, writes that “the present system of commercial surety bail should be simply
and totally abolished. . . . It is not so much that bondsmen are evil—although they sometimes
are—but rather that they serve no useful purpose.” American Bar Association, Criminal Justice
Standards, ch. 10, Pretrial Release, Standard 10–5.5, Compensated Sureties, 114–15 (1985),
refers to the commercial bond business as “tawdry” and discusses “the central evil of the
compensated surety system.” When Oregon considered reintroducing commercial bail, Judge
William Snouffer testified, “Bail bondsmen are a cancer on the body of criminal justice” (quoted
in Spurgeon Kennedy & D. Alan Henry, Commercial Surety Bail: Assessing Its Role in the
Pretrial Release and Detention Decision (1997)). Supreme Court Justice Harry Blackmun called
the commercial bail system “offensive” and “odorous.” See Schilb v. Kuebel, 404 U.S. 357
(1971).
8
In order to provide appropriate incentives, money bail is typically higher for the rich than
the poor. Thus, it is not a priori necessary that money bail should discriminate against the
poor, although in practice this does occur owing to nonlinearities and fixed costs in the bail
process. Assume that money bail is set so as to create equal FTA rates across income classes.
In such a case, there is no discrimination against the poor in the setting of bail. But if the bail
significant efforts were made, beginning in the 1960s, to develop alternatives

to money bail, and four states—Illinois, Kentucky, Oregon, and Wisconsin—
have outlawed commercial bail altogether.
In place of commercial bail, Illinois introduced the Illinois Ten Percent
Cash Bail or “deposit bond” system. In a deposit bond system, the defendant
is required to post with the court an amount up to 10 percent of the face
value of the bond. If the defendant fails to appear, the deposit may be lost
and the defendant held liable for the full value of the bond. If the defendant
appears for trial, the deposit is returned to the defendant, less a small service
fee in some cases.9 Some counties will also release defendants on unsecured
bonds. Unsecured bonds are equivalent to zero-percent deposit bonds. That
is, defendants released on an unsecured bond are liable for the full bail amount
if they fail to appear, but they need not post anything to be released.
The pretrial release system is designed to ensure that defendants appear
in court. It is often asserted that the commercial bail system discourages
appearance. In a key Supreme Court case, for example, Justice Douglas
argued that “the commercial bail system failed to provide an incentive to
the defendant to comply with the terms of the bond. Whether or not he
appeared at trial, the defendant was unable to recover the fee he had paid
to the bondsman. No refund is or was made by the professional surety to a
defendant for his routine compliance with the conditions of his bond.”10
Similarly, Jonathan Drimmer said, “Hiring a commercial bondsman re-
moves the incentive for the defendant to appear at trial.”11 John S. Goldkamp
and Michael R. Gottfredson suggest that the “use of the bondsman defeated
the rationale that defendants released on cash bail would have an incentive
to return,”12 and in their influential set of performance standards for pretrial
release, the National Association of Pretrial Service Agencies13 said that under
commercial bail, “the defendant has no financial incentive to return to
court.”14
In light of the persistent criticism that surety bail encourages failure to
appear, it is perhaps surprising that the data consistently indicate that defen-
dants released via surety bond have lower FTA rates than defendants released
under other methods. Part of this might be explained by selection—FTA
amounts necessary to ensure equal FTA rates are not linear in wealth, then such rates can
generate unequal rates of release across income classes.
9
National Association of Pretrial Service Agencies, Performance Standards and Goals for
Pretrial Release (2d ed. 1998).
10
Schilb v. Kuebel, 404 U.S. at 373.
11
Jonathan Drimmer, When Man Hunts Man: The Rights and Duties of Bounty Hunters in
the American Criminal Justice System, 33 Hous. L. Rev. 731, 742 (1996).
12
John S. Goldkamp & Michael R. Gottfredson, Policy Guidelines for Bail: An Experiment
in Court Reform 19 (1985).
13
See note 9 supra.
14
See also Thomas, ed., supra note 7, at 13. Because of this issue, Thomas calls the surety
system “irrational.”
bail jumping 97
rates, for example, may be higher for those defendants charged with minor
crimes—perhaps these defendants reason that police will not pursue a failure
to appear when the underlying crime is minor—and defendants charged with
minor crimes are more likely to be released on their own recognizance than
on surety release. A second reason, however, is that bond dealers, just like
other lenders, have numerous ways of creating appropriate incentives for
borrowers.
Most obviously, a defendant who skips town will owe the bond dealer the
entire amount of the bond. Defendants are often judgment proof, however,
so bond dealers ask defendants for collateral and family cosigners to the
bond (which is not done under the deposit bond system). If hardened criminals
do not fear the law, they may yet fear their mother’s wrath should the bond
dealer take possession of their mother’s home because they fail to show up
for trial. In order to make flight less likely, bond dealers will also sometimes
monitor their charges and require them to check in periodically. In addition,
bond dealers often remind defendants of their court dates and, perhaps more
important, remind the defendant’s mother of the son’s court date when the
mother is a cosigner on the bond.15
If a defendant does fail to appear, the bond dealer is granted some time,
typically 90–180 days, to recapture him before the bond dealer’s bond is
forfeited. Thus, bond dealers have a credible threat to pursue and rearrest
any defendant who flees. Bond dealers report that just to break even, 95
percent of their clients must show up in court.16 Thus, significant incentives
exist to pursue and return skips to justice.
Bond dealers and their agents have powerful legal rights over any defendant
who fails to appear, rights that exceed those of the public police. Bail en-
forcement agents, for example, have the right to break into a defendant’s
home without a warrant, make arrests using all necessary force including
deadly force if needed, temporarily imprison defendants, and pursue and
return a defendant across state lines without the necessity of entering into
an extradition process.17
At the time they write the bond, bond dealers prepare for the possibility
of flight by collecting information that may later prove useful. A typical
application for bond, for example, will contain information on the defendant’s
residence, employer, former employer, spouse, children (names and schools),
spouse’s employer, mother, father, automobile (description, tags, financing),
15
See Mary A. Toborg, Bail Bondsmen and Criminal Courts, 8 Just. Sys. J. 141, 156 (1983).
Bail jumping is itself a crime that may result in additional penalties.
16
Drimmer, supra note 11, at 793 (1996); Morgan Reynolds, Privatizing Probation and Parole,
in Entrepreneurial Economics: Bright Ideas from the Dismal Science 117, 128 (Alexander
Tabarrok ed. 2002).
17
Drimmer, supra note 11. See also Taylor v. Taintor, 83 U.S. (16 Wall.) 366 (1873).
union membership, previous arrests, and so on.18 In addition, bond dealers

have access to all kinds of public and private databases. Bob Burton,19 a
bounty hunter of some fame, for example, says that a major asset of any
bounty hunter is a list of friends who work at the telephone, gas, or electric
utility, the post office, or welfare agencies or in law enforcement.20
Bond dealers, however, recognize that what makes their pursuit of skips
most effective is the time they devote to the task. In contrast, public police
are often strained for resources, and the rearrest of defendants who fail to
show up at trial is usually given low precedence.
The flow of arrest warrants for FTA has overwhelmed many police de-
partments, so today many counties are faced with a massive stock of unserved
arrest warrants. Baltimore alone had 54,000 unserved arrest warrants as of
1999.21 In recent years, Cincinnati has had over 100,000 outstanding arrest
warrants stemming from failures to appear in court. One Cincinnati defendant
had 33 pending arrest warrants against him.22 In response to the overwhelming
number of arrest warrants, most of which will never be served because of
lack of manpower, some counties have turned to extreme measures such as
offering amnesty periods. Santa Clara County in California, for example, has
a backlog of 45,000 unserved criminal arrest warrants and in response has
advertised a hotline that defendants can use to schedule their own arrests.23
Although national figures are not available, it is clear that the problem of
outstanding arrest warrants is widespread. Texas, for example, is relatively
well off with only 132,000 outstanding felony and serious misdemeanor
warrants, but Florida has 323,000, and Massachusetts, as of 1997, had around
275,000.24 California has the largest backlog of arrest warrants in the nation.
The California Department of Corrections estimated that as of December
1998, there were more than 2.5 million unserved arrest warrants.25 Many of
these arrest warrants are for minor offenses, but tens of thousands are for
18
We thank Bryan Frank of Lexington National Insurance Corporation for discussion and
for sending us a typical application form.
19
Bob Burton, Bail Enforcer: The Advanced Bounty Hunter (1990).
20
Good bond dealers master the tricks of their trade. One bond dealer pointed out to us, for
example, that the first three digits in a social security number indicate in what state the number
was issued. This information can suggest that an applicant might be lying if he claims to have
been born in another state (many social security numbers are issued at birth or shortly thereafter),
and it may provide a lead for where a skipped defendant may have family or friends.
21
Francis X. Clines, Baltimore Gladly Breaks 10-Year Homicide Streak, N.Y. Times, January
3, 2001, at A11.
22
George Lecky, Police Name “200 Most Wanted,” Cincinnati Post, September 5, 1997, at
1A.
23
See Jane Prendergast, Warrant Amnesty Offered for 1 Day, Cincinnati Enquirer, November
19, 1999, for description of a similar program in Kenton County, Kentucky. See also Henry
K. Lee & Kenneth Howe, Plan to Clear Backlog of Warrants: Santa Clara County Offering
Amnesty to Some, S.F. Chron., January 12, 2000, at A15.
24
Howe and Hallissy, supra note 3.
25
Id.
bail jumping 99
people wanted for violent crimes, including more than 2,600 outstanding
homicide warrants.26 Kenneth Howe and Erin Hallissy report that “local, state
and federal law enforcement agencies have largely abandoned their job of
serving warrants in all but the most serious cases.” Explaining how this
situation came about, they write, “As arrests increased, jails became over-
crowded. To cope, judges, instead of locking up suspects, often released them
without bail with a promise to return for their next court date. For their part,
police, rather than arrest minor offenders, issued citations and then released
the suspects with the same expectation. When suspects failed to appear for
their court dates, judges issued bench warrants instructing police to take the
suspects into custody. But this caused the number of warrants to balloon,
and the police did not have the time or staff to serve them all.”27
III. The Matching Model with Multiple Treatments

Ideally, in a treatment evaluation we would like to identify two outcomes:
one if the individual is treated, YT, and one if no treatment is administered,
YNT. The effect of the treatment is then YT ⫺ YNT. But we cannot observe an
individual in both states of the world, making a direct computation of
YT ⫺ YNT impossible.28 All methods of evaluation, therefore, must make some
assumptions about “comparable” individuals. An intuitive method is to match
each treated individual with a statistically similar untreated individual and
compare differences in outcomes across a series of matches. Thus, two sta-
tistical doppelgangers would function as the same individual in different
treatments.
An important advantage of matching methods is that they do not require
assumptions about functional form. When the research question is about a
mean treatment effect, as it is here, matching methods also allow for an
economy of presentation because they focus attention on the question of
interest rather than on a long series of variables that are used only for control
purposes. Unfortunately, matching methods typically founder between a rock
and a hard place. The technique works best when individuals are matched
across many variables, but as the number of variables increases, the number
of distinct “types” increases exponentially, so the ability to find an exact
match falls dramatically.
In an important paper, Paul Rosenbaum and Donald Rubin go a long way
to surmounting this problem.29 They show that if matching on X is valid,
then so is matching on the probability of selection into a treatment condi-
tional on X. The multidimensional problem of matching on X is thus trans-
26
Id.
27
Id.
28
Rubin, Estimating Causal Effects of Treatments, supra note 4.
29
Paul R. Rosenbaum & Donald B. Rubin, The Central Role of the Propensity Score in
Observational Studies for Causal Effects, 70 Biometrika 41, 55 (1983).
formed into a single-dimension problem of matching on Pr(T p 1FX), where

T p 1 denotes treatment.30 The probability Pr(T p 1FX) is often called the
propensity score, or p-score.
The matching technique extends naturally to applications with multiple
treatments through the use of a multivalued propensity score with matching
on conditional probabilities.31 Assume that there are M mutually exclusive
treatments, and let the outcome in each state be denoted Y1, Y2, and so forth.
As before, we observe only a specific outcome but are interested in the
counterfactual: what would the outcome have been if this person had been
assigned to a different treatment? Rather than a single comparison, we are
now interested in a series of pairwise comparisons between treatments m
and l. The treatment effect on the treated is written
v0m,l p E(Y m ⫺ Y lFT p m) p E(Y mFT p m) ⫺ E(Y lFT p m), (1)
where v0m,l denotes the effect of treatment m rather than l.

Identification of (1) can occur under appropriate conditions, the most im-
portant being that treatment outcomes are independent of treatment selection
after conditioning on a vector of attributes, X (the conditional independence
assumption). Formally,
Y 1 … Y M ⊥ TFX p x. (2)
If this assumption is valid, we can use the conditional propensity score to

identify the treatment effect,32
v m,l p E(Y mFT p m) ⫺ E pmFml[E(Y lFp mFml (X), T p l)FT p m]. (3)
In practice, the conditional propensity score, p mFml (x), is computed indirectly
30
Matching methods are common among applied statisticians and natural scientists but have
only recently been analyzed and applied by econometricians and economists. Papers on the
econometric theory of matching include Heckman, Ichimura, & Todd, supra note 4; and Guido
W. Imbens, The Role of the Propensity Score in Estimating Dose-Response Functions (Technical
Working Paper No. 237, Nat’l Bur. Econ. Res. 1999). More applied work includes James J.
Heckman, Hidehiko Ichimura, & Petra E. Todd, Matching as an Econometric Evaluation Es-
timator: Evidence from Evaluating a Job Training Program, 64 Rev. Econ. Stud. 605, 654
(1997); Dehejia & Wahba, supra note 4; Michael Lechner, Programme Heterogeneity and
Propensity Score Matching: An Application to the Evaluation of Active Labour Market Policies
(Contributed Paper No. 647, Econ. Soc’y World Congress 2000). Our multitreatment application
is closest to that of Michael Lechner, Identification and Estimation of Causal Effects of Multiple
Treatments under the Conditional Independence Assumption (Discussion Paper No. 91, IZA
1999).
31
Lechner, Identification and Estimation of Causal Effects, supra note 30; Imbens, supra
note 30.
32
Lechner, Identification and Estimation of Causal Effects, supra note 30.
bail jumping 101
from the marginal probabilities p l (x) and p m (x) estimated from a discrete-
choice model. In this case,
p m (x)
E[ p mFml (x)Fp l (x), p m (x)] p E [ p l (x) ⫹ p m (x) ]
Fp l (x), p m (x) p p mFml (x).
(4)
We use an ordered probit model (see further below) to generate propensity
scores.
It is important to emphasize that the propensity scores are not of direct
interest but rather are the metric by which members of the treated group are
matched to members of the “untreated” group (“differently” treated in our
context). After matching, and given the conditional independence assumption,
the treated and untreated groups can be analyzed as if treatment had been
assigned randomly. Thus, differences in mean FTA rates across matched
samples are estimates of the effect of treatment.
Less formally, matching on propensity scores can be understood as a
pragmatic method for balancing the covariates of the sample across the dif-
ferent treatments.33 Note that the covariates that we care most about balancing
are those that affect the treatment outcome. Assume, for example, that X
influences treatment selection but does not independently influence treatment
outcome. If the goal of the selection model were to consistently estimate the
causes of treatment selection, we would want to include X in the model, but
it is not necessarily desirable to include it when the purpose is to create a
metric for use in matching.34 A simple example occurs when X predicts
treatment exactly. Inclusion of X would defeat the goal of matching because
all propensity scores would be either zero or one. Similarly, we will include
model variables in the propensity score that may affect the treatment outcome
even if they do not casually affect treatment selection.
IV. Data and Descriptive Statistics

We use a data set compiled by the U.S. Department of Justice’s Bureau
of Justice Statistics called State Court Processing Statistics, for 1990, 1992,
1994, and 1996 (Inter-university Consortium for Political and Social Research
[ICPSR] study 2038). We supplement these data with an earlier version of
the same collection, the National Pretrial Reporting Program, for 1988–89
(ICPSR study 9508). The data are a random sample of 1 month of felony
filings from approximately 40 jurisdictions, where the sample was designed
to represent the 75 most populous U.S. counties. The data contain detailed
information on arrest charges, criminal background of the defendant (for
33
Dehejia & Wahba, supra note 4.
34
Boris Augurzky & Christoph M. Schmidt, The Propensity Score: A Means to an End
(Discussion Paper No. 271, IZA 2001).
example, number of prior arrests), sex and age of the defendant,35 release
type (surety, cash bond, own recognizance, and so on), rearrest charges for
those rearrested, whether the defendant failed to appear, and whether the
defendant was still at large after 1 year, among other categories.
In addition to the main release types, there are minor variations. Some
counties, for example, release on an unsecured bond for which the defendant
pays no money to the court but is liable for the bail amount should he fail
to appear. Because the incentive effects are very similar, we include unsecured
bonds in the deposit bond category.36 Instead of a pure cash bond, it is
sometimes possible to put up property as collateral. Since property bonds
are rare (588 observations in our data, less than 2 percent of all releases),
we drop them from the analysis.37 Finally, some counties may occasionally
use some form of supervised release. In the first year of our data set, su-
pervised release is included in the own-recognizance category. Supervised
release often means something as simple as a weekly telephone check-in, so
including these with own recognizance is reasonable. Supervised release is
not a standard term, however, and other forms, such as mandatory daily
attendance of a drug treatment program, are likely to be more binding. To
maintain comparability across years, we follow the practice established in
the first year of the data set by classifying supervised release with own
recognizance. Because supervised release is more binding than pure own
recognizance, this can only lower FTA rates and other results in the own-
recognizance sample, thus biasing our results away from finding significant
differences among treatments.38
In Table 1, the mean FTA rates for release categories are along the main
diagonal, with the number of observations in square brackets. The preliminary
analysis suggests that FTA rates are lower under surety bond release than
under most other types of release. Off-diagonal elements are the difference
between the FTA rate for the row category and the FTA rate for the column
category. The FTA rate for those released under surety bond is 17 percent.
Compared with surety release, the FTA rate is 3 percentage points higher
under cash bonds, 4 percentage points higher under deposit bonds, and 9
percentage points higher under own recognizance (all these differences are
35
The State Court Processing Statistics data are more complete and better organized than
the National Pretrial Reporting Program data. The former, for example, include information
on the race of the defendant that the latter do not.
36
We drop observations missing data on the bail amount.
37
Another reason to drop property bonds is that it is difficult to compare the bail for these
releases to other release types. A defendant, for example, may put up a $250,000 house as
collateral for $25,000 in bail. Although we know the bail amount, we do not know the value
of the collateral property other than that it must, by law in many cases, be higher than the
value of the bail amount. A cash or surety bond, therefore, is not equivalent to a property bond
for the same bail amount.
38
We find similar results by restricting the data set to the years in which supervised release
is given a distinct category.
bail jumping 103
TABLE 1
Mean Failure-to-Appear Rates by Release Category, 1988–96
Own Deposit Cash Surety Emergency

Recognizance Bond Bond Bond Release
Own recognizance 26 [20,944] 5** 6** 9** ⫺19**
Deposit bond 21 [3,605] 1 4** ⫺23**
Cash bond 20 [2,482] 3** ⫺25**
Surety bond 17 [9,198] ⫺28**
Emergency release 45 [584]
Note.—Mean failure-to-appear (FTA) rates (in %) for release categories, rounded to the nearest integer,
are along the main diagonal, with the number of observations in square brackets. Off-diagonal elements
are the difference between the mean FTA rate for the row category and the mean FTA rate for the column
category.
** Statistically significant at the greater than 1% level.
statistically significant at greater than the 1 percent level). Put slightly dif-
ferently, compared with surety release, the FTA rate is approximately 18
percent higher under cash bond, 33 percent higher under deposit bond, and
more than 50 percent higher under own recognizance.
Table 1 also presents some information on emergency release. Emergency
releasees are defendants who are released solely because of a court order to
relieve prison overcrowding. Emergency release is not a treatment—the treat-
ment is own recognizance—but rather an indication of what happens when
neither judges nor bond dealers play their usual role in selecting defendants
to be released.39 One would expect that relative to those released under other
categories, these defendants are likely to be accused of the most serious
crimes, have the highest probability of being found guilty, and have the
fewest community ties. In addition, these defendants have neither monetary
incentive nor the threat of being recaptured by a bounty hunter to induce
them to return to court. As a result, a whopping 45 percent of the defendants
who are given emergency release fail to appear for trial. The large differences
between the FTA rates of those released on emergency release and every
other category indicate that substantial and successful selection occurs in the
decision to release. Emergency release is thus of some special interest, al-
though not directly related to the focus of this paper.
Although the preliminary data analysis is suggestive, the difference-in-
means analysis could confound effects due to treatment with effects due to
selection on, for example, defendant characteristics such as the alleged crime.
39
Even under emergency release, some selection can occur. Judges and jailers, for example,
could order that more inmates be paroled to make room for the most potentially dangerous
accused defendants, inmates could be shipped out of state, or the court order could be (tem-
porarily) ignored. The costs of selection, however, clearly rise substantially when jail space is
tightly constrained.
V. Results
A. Propensity Scores from Ordered Probit

We generate propensity scores for matching using an ordered probit model.
By law, judges must release defendants on the least restrictive conditions
they believe are compatible with ensuring appearance at trial. Own recog-
nizance, the least restrictive form of release, is our first category, followed
by release on deposit bond. Although defendants released on deposit bond
must put up some cash, which they will forfeit if they fail to appear, the
amount is typically less than $500.40 Few people are ever held because of a
failure to raise cash for a deposit bond. Defendants who were offered financial
release (but not a deposit bond) and who paid their bonds in cash are the
third category of release. Cash bond is more expensive than a deposit bond
but does not involve the monitoring of sureties. Defendants released via
surety bond are the fourth category. Although the Constitution guarantees
that excessive bail shall not be required, it does not require that bail should
always be set low enough for a defendant to be able to afford release. Indeed,
judges sometimes set bail in the expectation (and hope) that the defendant
will not be able to raise bail. Thus, we include defendants held on bail or
detained without bail as the final, most restrictive category, not released.
Emergency releases are also included in the final category because, had it
not been for the emergency, these individuals would have not have been
released. From the ordered probit, we generate conditional propensity scores
for each possible pairwise comparison.41
Variables in the ordered probit specification include individual-specific
indicators that denote whether the defendant has been accused of murder,
rape, robbery, assault, other violent crime, burglary, theft, other property
offense, drug trafficking, other drug-related offense, or driving-related offense
(with misdemeanors and other crimes in the constant). We also include var-
iables for past experience with the criminal justice system. Three binary
variables are set equal to one if the defendant had some active criminal
justice status at the time of the arrest (for example, was on parole or pro-
bation), had prior felony arrests, or failed appear at trial in the past. The
defendant’s sex and age are also included. Note that these variables are
exactly the sorts of variables that judges use to make treatment selection
40
The median deposit bond amount is $5,000, and releasees typically must deposit 10 percent
or less of the bond amount.
41
We have also estimated the results using a multivariate logit model. The results are sub-
stantively similar (on the ordered probit model, see, for example, William H. Greene, Econ-
ometric Analysis (4th ed. 2000)).
bail jumping 105
decisions.42 Other, nonindividual variables include the police clearance rate,

defined as the number of arrests divided by the number of crimes per county.
The clearance rate provides a crude measure of police availability that may
affect FTA rates. County and year effects are included in the selection equa-
tion (county 29 and 1988 are excluded to prevent multicollinearity).43 The
results of the ordered probit estimation are presented in Appendix Table A1.
B. Matching Quality
A match is defined as the pair of observations with the smallest difference
in propensity scores so long as the difference is less than a predefined caliper.
If a match cannot be made within the caliper distance, the observations are
dropped. We use matching with replacement, so the order of matching is
irrelevant, and every untreated observation is compared against every treated
observation.44
The match quality is good, as we match large proportions of the sample
despite using a caliper of only .0001.45 Figure 1A presents a box-and-whiskers
plot of the propensity scores for each treatment category (including the “treat-
ment” of not released) conditional on the actual treatment. The leftmost part
of the graph, for example, gives the box-and-whiskers plot for the propensity
of being in the own-recognizance, deposit, cash, surety, and not-released
treatments for all defendants in the own-recognizance treatment.46
42
Ayres & Waldfogel, supra note 6, identifies eight characteristics that judges may consider
in setting bail: (1) the nature and circumstances of the offense (if relevant), (2) the evidence
against the defendant, (3) the defendant’s prior criminal record, (4) the defendant’s prior FTA
record, (5) the defendant’s family ties, (6) the defendant’s employment record, (7) the de-
fendant’s financial resources, and (8) the defendant’s community ties. Although Ayers and
Waldfogel’s study deals only with Connecticut, the criteria are similar in other states.
43
The use of county effects in the selection equation is noteworthy because it implies that
matching will occur with “quasi”-fixed effects. A true fixed-effects estimator would require
that comparable observations come from within the same county. The matching estimator takes
into account county effects when seeking a match but does not insist that every match must
be within county. In particular, some counties do not release on deposit bond, and others do
not release on surety bond. A fixed-effects estimator would not use information from these
counties in estimating the effect of the deposit and surety treatments. The matching estimator
will use information from these counties if matching is strong on other variables. A pure fixed-
effects estimator may also be important, however, and in the working version of this paper,
Eric Helland & Alexander Tabarrok, Public versus Private Law Enforcement: Evidence from
Bail Jumping (Working paper, George Mason Univ. 2003), we pursue this alternative approach.
Results are consistent with those discussed here.
44
Dehejia & Wahba, supra note 4, finds that matching with replacement is considerably
superior to matching with nonreplacement.
45
When matching on variables with fewer observations, such as fugitive rates conditional
on failure to appear as we do below, we match using a caliper of .001. The caliper size makes
little difference to the results.
46
In a box-and-whiskers plot, the box contains the interquartile range (IQR): the observations
between the 75th percentile (the top of the box) and the 25th percentile (the bottom of the
box). The horizontal line toward the center of each box is the median observation. The whiskers
are the so-called adjacent values that extend from the largest observation less than or equal
Figure 1.—A, p-score distribution for each release type conditional on actual release (the
order within type is own recognizance, deposit, cash, surety, not released); B, pairwise p-score
distributions for own recognizance versus surety.
bail jumping 107
Figure 1B gives the box-and-whiskers plot for the pairwise (conditional)

probabilities for the own versus surety comparison. The “Pr. Own” and “Pr.
Surety” arrows indicate that we can find comparable observations, statistical
doppelgangers, for individuals released under either treatment. Many of the
defendants released on surety bond, for example, were as likely to have been
released on their own recognizance (third box from the left) as those who
actually were released on their own recognizance (first box from the left).
Similarly, many of the defendants who were released on their own recog-
nizance were as likely to have been released on surety bond (second box
from the left) as those who actually were released on surety bond (fourth
box from the left). Note that it is important that the boxes overlap across
treatments, not that they overlap within treatments—that is, the fact that in
Figure 1A the propensity to be in the deposit bond treatment is everywhere
lower than the propensity to be in the own-recognizance treatment simply
reflects the fact that the deposit bond treatment is a low-probability event.
More important is that the deposit bond treatment is a low-probability event
regardless of actual treatment—we can thus find comparable observations
across the treatments. Alternatively stated, the overlap in the boxes across
treatments indicates that random factors play a large role in treatment se-
lection, thus aiding our effort to find true comparable observations.47
Although we can find comparable observations across the release treat-
ments, we cannot find good comparable observations for those who were
not released. Indeed, the Figure 1A box-and-whiskers plot of the propensity
not to be released among those who in fact were not released does not overlap
at all with the propensity not to be released for those who were released.
Defendants who are not released differ greatly from released defendants.48
(This is consistent with the very high FTA rates that we found for emergency
releasees in Table 1.) The fact that the model is capable of finding large
selection effects if they exist, as they apparently do for those not released,
bolsters the finding that selection on observable characteristics is not overly
strong among the release treatments.
to the 75th percentile plus 1.5 # IQR and the smallest observation greater than or equal to
the 25th percentile minus 1.5 # IQR. Points outside the box and whiskers are called extreme
values or outside points and for clarity are not plotted in this graph. In this plot, the width of
the box is proportional to the square root of the number of observations in that category.
47
Another interesting aspect of the box-and-whiskers plot is that it suggests that almost
everyone can be released on their own recognizance, even those who might in another time
and place be released only with high bail. Thirty percent of released defendants accused of
murder, for example, were released on their own recognizance.
48
It is possible to find defendants who were released who might not have been released—
thus, the data are consistent with the adage that it is better to let 10 guilty men go free than
jail one innocent man.
TABLE 2
Treatment Effects of Row versus Column Release Category on Failure-to-
Appear Rates using Matched Samples, 1988–96
Own Deposit Cash Surety

Recognizance Bond Bond Bond
Own recognizance 26 3.2** (1.0; 1.1) 4.8** (1.1; 1.2) 6.5** (.78; .78)
Deposit bond ⫺3.1** (1.1; 1.2) 21 4.1** (1.5; 1.6) 3.1** (1.1; 1.3)
Cash bond ⫺5.8** (1.3; 1.6) ⫺1.5 (1.6; 2.0) 20 1.8; 2.0 (1.4; 1.8)
Surety bond ⫺7.3** (.78; .89) ⫺3.9** (1.1; 1.2) 1.7 (1.3; 1.4) 17
Note.—Mean failure-to-appear rates (in %) for release categories for the full sample are along the main
diagonal. Off-diagonal elements are the estimated treatment effects of the row category versus the column
category. Standard errors are in parentheses—the first standard error assumes that the p-score is estimated
with certainty; the second uses bootstrapping to estimate the standard error including uncertainty of the
p-score. Matching caliper p .0001.
** Statistically significant at the greater than 1% level (two sided).
C. Estimated Treatment Effects: Failure to Appear

In Table 2, the row variable denotes the treated variable and the column
the untreated variable.49 For reference, the main diagonal includes the mean
FTA rate in that category from the full sample.50 Reading across the surety
bond row, for example, we see the estimated difference in FTA rates caused
by the surety treatment relative to the column treatment—that is, the estimate
of the effect of treatment on the treated. The matching estimator suggests
that similar individuals are 7.3 percentage points, or 28 percent, less likely
to fail to appear when released on surety bond than when released on their
own recognizance. Similar individuals are also 3.9 percentage points, or 18
percent, less likely to fail to appear when released on surety bond than when
released on deposit bond. The estimated treatment effect for those on surety
bonds versus cash is small and not statistically significant.51
49
Two standard errors are presented in Table 2. The first takes into account uncertainty in
the matched samples but assumes that the propensity score is known with certainty. The second
estimate is a bootstrapped standard error that takes into account uncertainty that propagates
from the estimation of the propensity score. The “regular” and bootstrapped standard errors
are close, with the bootstrapped errors being approximately 8–20 percent higher. All the sta-
tistically significant results are significant at greater than the 1 percent level using either standard
error. Since the estimation of the propensity score adds very little uncertainty to the matching
estimators and because calculating bootstrapped errors is very time and resource intensive, we
present only the regular standard errors in future results and leave adjustments to the reader.
The bootstrapped errors were calculated using 100 replications of the model. The procedure
took over 48 hours on a reasonably fast Pentium computer.
50
The mean FTA rate for the full sample is included as rough guide to absolute effects.
Note, however, that the matched sample is usually smaller than the full sample, so the mean
FTA rate for the matched and full samples can be slightly different.
51
As a test of matching quality, we also ran a linear regression on the matched samples that
included surety bond and all the variables in Table 3. The results are similar, as they should
be if the matched samples divide other covariates as if they were assigned randomly. The
coefficient on surety bond in the surety versus own recognizance regression, for example, is
bail jumping 109
Unlike Table 1, both the top and bottom halves of Table 2 are filled in;
this is because the estimate of the treatment on the treated is conceptually
different from the estimate of the treatment on the untreated (differently
treated). For example, the effect of the surety treatment relative to own
recognizance for those who were released on surety bond is not necessarily
the exact opposite of the effect of own recognizance relative to surety bond
on those who were released on their own recognizance. As it happens, how-
ever, our estimates of these effects are similar. The estimate of the effect of
own recognizance relative to surety on those who were released on their own
recognizance, for example, is 6.5 percentage points, similar in size but op-
posite in sign to the ⫺7.3 surety effect relative to own recognizance of those
who were released on surety bond. The similarities across diagonals suggest
that either (or both) treatment selection or treatment effect does not interact
strongly with defendant characteristics. One possible exception is that the
deposit bond treatment relative to cash is estimated at 4.1 percentage points,
while the cash bond treatment relative to deposit is estimated at ⫺1.5 per-
centage points.
D. Estimated Treatment Effects: The Fugitive

A surprisingly large number of felony defendants who fail to appear remain
at large after 1 year, approximately 30 percent. Alternatively stated, some
7 percent of all released felony defendants skip town and are not brought
back to justice within 1 year. Those who remain at large more than 1 year
are called fugitives.
The surety treatment differs most from other treatments when a defendant
purposively skips town, because this is when bounty hunters enter the pic-
ture.52 If the surety treatment works, therefore, we should see it most clearly
in the apprehension of fugitives. Given that a defendant fails to appear, we
ask what the probability is that the defendant is not brought to justice within
1 year and how this varies with release type. It is important to note that once
a defendant has decided to abscond, there is no reason why anything other
than the different effectiveness of public police and bail enforcement agents
should have a systematic effect on the probability of being recaptured.
Table 3 provides strong evidence that bounty hunters are highly effective
at recapturing defendants who attempt to flee justice—considerably more so
than the public police. The main diagonal of Table 3 contains the mean
fugitive rate conditional on FTA along with the number of observations in
⫺6.5, which is within 1 standard deviation of the ⫺7.3 matching estimate. We do a more
detailed comparison of linear regression and matching results further below.
52
We use the term “bounty hunter” or “bail enforcement agent” to refer to private pursuers
of felony defendants. Bond dealers typically pursue their own skips. Literal bounty hunters
are typically not called in unless the skip is thought to have crossed state or international lines.
Services like Wanted Alert (http://www.wantedalert.com) regularly post ads in USA Today that
list fugitives and their bounties.
TABLE 3
Treatment Effect of Row versus Column Release Category on the Fugitive Rate
using Matched Samples, Conditional on Failure to Appear, 1988–96
Own Deposit Cash Surety

Recognizance Bond Bond Bond
Own recognizance 32 [5,440] ⫺3** (2.6) ⫺4.9** (2.9) 9.4** (2.1)
Deposit bond ⫺.2 (2.6) 33 [766] ⫺6.2 (4.1) 12.1** (2.7)
Cash bond 11.9** (3.0) ⫺3.8 (4.4) 40 [506] 18.6** (3.7)
Surety bond ⫺17** (2.0) ⫺15.5** (2.9) ⫺25.6** (4.2) 21 [1,537]
Note.—Mean fugitive rates (in %), defined as failures to appear that last longer than a year, for release
categories for the full sample are along the main diagonal, with the number of observations in that category
conditional on a failure to appear in square brackets. Off-diagonal elements are the difference between the
mean fugitive rate for the row category and the mean fugitive rate for the column category estimated using
matching. Standard errors are in parentheses. Matching caliper p .001.
each category. The estimated treatment effects for the row versus column
variables are shown in the off-diagonals with standard errors in parentheses.
The probability of remaining at large for more than a year conditional on
an initial FTA is much lower for those released on surety bond. The surety
treatment results in a fugitive rate that is lower by 17, 15.5, and 25.6 per-
centage points compared with the own-recognizance, deposit bond, and cash
bond treatments, respectively. In percentage terms, the fugitive rates under
surety release are 53, 47, and 64 percent lower than the fugitive rates under
own recognizance, deposit bond, and cash bond, respectively. Similarly, the
own recognizance, deposit, and cash bond treatments result in fugitive rates
that are 29, 47, and 47 percent higher than under the surety treatment.
There are also some interesting nonsurety effects in Table 3. Note that the
fugitive rate conditional on an FTA is higher for cash bond than for release
on own recognizance. Earlier (see Table 2) we had found that the FTA rate
was lower for cash bond than for release on own recognizance. This suggests
that defendants on cash bond are less likely to fail to appear than those
released on their own recognizance, but if they do fail to appear, they are
less likely to be recaptured. The result is pleasingly intuitive. A defendant
released on his own recognizance has little to lose from failing to appear
and thus may fail to appear for trivial reasons. But a defendant released on
cash bond has much to lose if he fails to appear, and thus those who do fail
to appear do so with the goal of not being recaptured.
The propensity score method can be very informative about the entire
distribution of treatment effects. In Figure 2, we graph smoothed (running-
mean) FTA and fugitive rates against surety p-scores for the own-
recognizance and surety treatments (conditional on being in either the surety
or own-recognizance treatment). (We omit graphs for the other treatment
comparisons for brevity.) The two downward-sloping, thinner curves graph
smoothed FTA rates against the p-scores for those defendants released on
bail jumping 111
Figure 2.—Failure-to-appear and fugitive rates by own recognizance versus surety treatment
plotted against p-scores.
their own recognizance or surety bond. The slope of each line indicates the
direction and strength of the effect of observable characteristics on selection
in that treatment. The difference between the own-recognizance and surety
lines at any given propensity score is an estimate of the treatment effect,
controlling for observable characteristics. The difference is roughly constant,
which indicates that despite some mild selection, the treatment effect is
roughly independent of observable characteristics.
For both the own-recognizance and surety treatments, FTA rates decrease
as the propensity for being in the surety treatment increases. That is, FTA
rates decrease as observable characteristics move in the direction of predicting
surety release. The decline is gentle; moving from a near-zero propensity to
a near-one propensity reduces the FTA rate by approximately 5 percentage
points. The effect is sensible if we recall that many FTAs are short term—
the defendant forgets the trial date or has another pressing engagement. These
sorts of FTAs are likely to be more common for defendants with observable
characteristics that predict low p-scores because judges release most defen-
dants on their own recognizance and reserve surety release for defendants
accused of more serious crimes. Few people will forget to show up for their
murder trial, but some may do so if the trial involves a driving offense. At
the same time, however, we expect that defendants accused of more serious
crimes—who have more to lose from being found guilty—are more likely
to purposively abscond. If this is correct, we ought to see a positive correlation

between the surety propensity score and the fugitive rate conditional on
having failed to appear.
The two upward-sloping, thicker lines plot smoothed fugitive rates against
the surety propensity score. As before, the slope of the plots gives the di-
rection and strength of effects caused by selection on observable character-
istics, and the vertical difference is the treatment effect for any given pro-
pensity score. As observable characteristics move in the direction of a greater
propensity to be selected for surety release, the fugitive rate increases. It is
interesting to note that the effect of selection on defendants released on surety
bond is less than that on defendants released on their own recognizance (that
is, the “slope” of the plot is less). This suggests that the surety treatment
works well even for those defendants whose observable characteristics would
predict higher FTA rates.
We examine the issue of unobservable characteristics at length below, but
since selection by observable characteristics has little influence on fugitive
rates, Figure 2 already suggests that observables would have to be very
different from observables in order to greatly affect the results.
E. Kaplan-Meier Estimation of Failure-to-Appear Duration

The higher rate of recapture for those released on surety bond compared
with other release types can be well illustrated with a survival function. For
a subset of our data, just over 7,000 observations, we have information on
the time from the failure to appear until recapture (return to the court). A
survival function graphs the percentage of observations that survive at each
time period. We estimate a survival function for each release type using the
nonparametric Kaplan-Meier estimator. Typically, the Kaplan-Meier esti-
mator is used only for preliminary analysis and is then followed by a para-
metric or semiparametric model. Although parametric and semiparametric
models allow for covariates, they require sometimes tenuous assumptions
about functional form. Instead, we follow our earlier approach of creating
matched samples. Thus, using the same procedure, we create three matched
samples: surety versus own recognizance, surety versus deposit, and surety
versus cash. We then compare the survival function across each matched
sample. The matching procedure ensures that covariates are balanced across
the matched samples, so it is not necessary to include additional controls for
covariates.
Figure 3 presents the survival functions. In each case, the survival function
for those released on surety bond is markedly lower than that for those
released on their own recognizance, deposit bond, or cash bond. The ability
of bail enforcement agents relative to police to recapture defendants who
bail jumping 113
Figure 3.—Kaplan-Meier survival function for defendants on surety bond versus those
released on cash bond or deposit bond or released on their own recognizance—using matched
samples.
skip bail is evident within a week of the failure to appear.53 By 200 days,
the surety survival rate is some 20–30 percentage points, or 50 percent, lower
than the survival rate for those out on cash bond, deposit bond, or their own
recognizance; that is, the probability of being recaptured is some 50 percent
higher for those released on surety bond relative to other releases. (Note that
there are three surety bond survival functions, one for each comparison group,
but these are nearly identical.)
A log-rank test confirms Figure 3; we can easily reject the null hypothesis
of equality of the survivor functions—defendants released on surety bond
are much more likely to be recaptured (that is, less likely to remain at large,
or “survive”) than those released on their own recognizance, deposit bond,
or cash bond.54
53
A number of estimates have been made that bounty hunters take into custody between
25,000 and 35,000 fugitives a year, depending on the year (see various sources in Drimmer,
supra note 11; and also W. P. Barr, letter to Charles T. Canady on the Bounty Hunter Re-
sponsibility Act, NABIC Bull., March 2000). These figures are consistent with a recapture rate
of over 95 percent and are consistent with the number of fugitives on surety bond. It appears,
therefore, that almost all fugitives on surety bond are recaptured by bail enforcement agents
and not by the police. Bounty hunters, however, will sometimes track down defendants and
then tip police as to their whereabouts, so police will sometimes be involved in some aspects
of recapture.
54
The exact results of the log-rank test and similar results matching on propensity score and
bail can be found in Helland & Tabarrok, supra note 43.
TABLE 4
Effect of Alternative Treatment versus Surety Bond on Failure-to-Appear and
Fugitive Rates (Conditional on Failure to Appear), 1988–96
Own Recognizance
versus Deposit versus Cash versus
Surety Bond Surety Bond Surety Bond
Treatment effect on failure-to-appear rates ⫹7.8** (1.6) ⫹6.2** (1.8) ⫺1.6 (4.4)
Treatment effect on fugitive rates ⫹14.8** (2.3) ⫹19.8** (2.9) ⫹35.7** (8.0)
Note.—Individuals from states that have banned surety bonds are matched with similar individuals
released on surety bond. Standard errors are in parentheses. The matching caliper is .0001.
F. Comparison with Counties in States That Have

Banned Commercial Bail
Some states have banned commercial bail. It seems plausible that matching
can find two individuals who are comparable but for the fact that one in-
dividual could not have been assigned surety bail while the other could and
was assigned surety bail. Comparing these individuals gives us a measure
of what would happen if a county lifted its ban on commercial bail.55
Table 4 demonstrates that states that ban commercial bail pay a high price.
We estimate that FTA rates are 7–8 percentage points, or approximately 30
percent, higher for individuals released on deposit or own recognizance than
if the same individuals were released on surety bond.56 As before, we find
that cash bond is about as effective as surety bond at controlling FTA rates.
The fugitive rate conditional on FTA is much higher under own recognizance,
deposit, or cash release than under surety—higher by some 15, 20, and 36
percentage points, or 78, 85, and 93 percent, respectively—figures even larger
than we found earlier.
VI. Looking for Unobservable Variables

Matching is a powerful and flexible tool, but it is not a research design
that magically guarantees the identification of causal effects. In this section,
we test for robustness and attempt to rule out the potentially confounding
effects of unobservable characteristics. We focus on two identification strat-
55
Since we are interested in the cross-county variation, the propensity scores for these tests
were generated from an ordered probit that did not include county fixed effects but was
otherwise identical to that used earlier.
56
Note that in Table 4, we examine the treatment effect of own recognizance, deposit, and
cash relative to surety because this is the relevant comparison when considering the experiment
of lifting the ban on commercial bail. As noted earlier, the treatment effect on the treated and
untreated groups are similar, so we could also have examined the surety treatment effect relative
to the alternative release types.
bail jumping 115
TABLE 5
Mean Rearrest Rates by Release
Category, 1988–96
Rate (%) N
Own recognizance 14.9 20,945
Deposit bond 13.3 3,605
Cash bond 14 2,482
Surety bond 12 9,202
egies; a number of alternative strategies, described briefly below, are devel-

oped in the working paper.
Our first identification strategy takes advantage of the fact that some 14
percent of defendants out on pretrial release are arrested for another crime
before they are sentenced for the first crime. It is plausible that the probability
of being rearrested is positively correlated with the probability of becoming
a fugitive. Assume, for example, that guilty defendants are less likely to
show up for trial than innocent defendants and that innocent defendants are
less likely to be rearrested than guilty defendants. There is good evidence
for some such assumption because in the raw data, defendants who are never
rearrested have an FTA rate of 11 percent, but defendants who are rearrested
for another crime have an FTA rate of 43 percent.
If rearrest is positively correlated with the probability of becoming a fu-
gitive and if treatment does not influence rearrest rates, then rearrest rates
by treatment will track unobserved characteristics. Table 5 provides evidence
for the second clause—in the raw data, there is very little variation in rearrest
rates across treatment categories.57 Thus, Table 6 (matching on propensity
score and bail) presents faux “treatment effects” for the effect of various
release types on rearrest rates. We emphasize that our hypothesis is that
treatment does not influence rearrest—the faux treatment effects, therefore,
are indications of the influence of unobserved variables.
In Table 6, the surety versus own recognizance and surety versus deposit
comparisons show positive but very small and statistically insignificant ef-
fects, which suggests that unobserved variables have little influence on FTA
and fugitive rates across these comparisons. The surety versus cash bond
comparison suggests that the surety treatment increases rearrest rates by 4.5
percentage points, which implies that unobserved variables operate in a di-
rection that offsets the true treatment effect of surety on FTA and fugitive
57
In the raw data, there appears to be a slight decrease in rearrest rates for those released
on commercial bail. Although the rearrest of a defendant is not usually grounds for the forfeiture
of the bond dealer’s bond, bond dealers do monitor their charges, and such monitoring might
reduce rearrest rates. Bond dealers might be also be able to select defendants who are unlikely
to flee and thus also unlikely to be rearrested. Once we control for observable characteristics,
however, the slight decrease in arrest rates for those on commercial bail disappears and in
some cases reverses (see Table 6).
TABLE 6
Effect of Surety Treatment Effect versus other Release Types on Rearrest
Rates using Samples Matched on p-Score and Bail, 1988–96
Surety versus Surety versus Surety versus

Own Recognizance Deposit Bond Cash Bond
Surety bond .7 (.6) .58 (1.0) 4.5** (1.3)
Matched observations 14,925 9,740 7,064
Note.—The matching caliper is .001.
rates. Recall from Table 2 that we found that FTA rates were slightly higher
under surety than under the cash bond treatment. The evidence from rearrest
rates suggests that unobservable characteristics may be responsible for part
of this and that the true treatment effect is somewhat lower. Similarly, al-
though we found large negative effects on fugitive rates from the surety
treatment (relative to cash treament), the evidence suggests that, if anything,
the true treatment effects are even more negative.58
The rearrest data allow for another interesting comparison. For a small
subset of our data, 1,331 observations from 1988 and 1990, we know the
rerelease type for those individuals who are arrested and released on a second
charge. We do not know whether the individual failed to appear on the second
charge, which is why we do not have repeated observations. Nevertheless,
the second arrest and release data may be revealing.
Suppose that the initial release is own recognizance and the second release
is via surety bond. By monitoring and possibly recapturing the defendant if
he skips on the second trial, bail bondsmen and their agents create a positive
externality with respect to fugitive rates on the first trial. This potential
externality means that we need not compare own-recognizance to surety
releases to measure a surety treatment effect. Instead, we can compare de-
fendants released on their own recognizance with other defendants released
on their own recognizance in their first release and on surety bond in their
second release. Similarly, we can compare fugitive rates on the first trial for
defendants whose first and second releases were own recognizance and own
recognizance with those whose first and second releases were own recognizance
and surety bond. With this comparison, we control for selection effects on the
first release.
The unconditional fugitive rate of defendants who are released on their
58
Since we find that rearrest rates vary little by treatment category, we should also find that
treatment effects measured in the rearrest sample, that is, using only those defendants who
were subsequently arrested for a second crime, should be similar to those found in the one-
arrest sample. We have run these matching tests on propensity score and bail and do find
similar results, which we omit for brevity.
bail jumping 117
TABLE 7
Unconditional Fugitive Rates by Arrest-
Rearrest Category, 1988, 1990
Category Rate
1. Own and not rearrested 8.48 [17,828]
2. Own-own 8.04 [191]
3. Own-surety 1.49 [134]
4. t-test (row 1 ⫺ row 3) 2.9; p(1 1 3 p .0019)
5. t-test (row 2 ⫺ row 3) 2.6; p(2 1 3 p .0047)
Note.—Own-own indicates first release on own recognizance

and second release on own recognizance. Own-surety indicates
first release on own recognizance and second release on surety
bond.
own recognizance and not rearrested is 8.48 percent.59 The fugitive rate of
defendants who are released on their own recognizance and who are rearrested
and then released again on their own recognizance is almost identical, 8.04
percent. But the fugitive rate for those defendants initially released on their
own recognizance but then rearrested and rereleased on surety bond is just
1.9 percent. The difference between the own-recognizance and the own-
recognizance⫹surety fugitive rate is statistically significant at the greater
than 1 percent level. The difference between the own-recognizance⫹own-
recognizance and own-recognizance⫹surety rate, which controls for rearrest,
is also statistically significant at the greater than 1 percent level. Table 7
summarizes.
In the working paper,60 we supplement the above analysis in a variety of
ways to control for county effects, individual effects observed by judges but
unobserved by us, and pure unobserved effects of a very general nature. 61
Most generally, the cream that judges skim are released on their own re-
cognizance and deposit bond, while the skim are released on cash or surety
bond. Consistent with this, observable selection effects on fugitive rates are
positive, and the evidence from a variety of independent tests suggests that
unobservable characteristics are not biasing our results upward. Taken to-
59
Earlier we focused on fugitive rates conditional on having FTA. We focus on unconditional
fugitive rates here because we have fewer observations. We have data on rearrest and rerelease
type for 1988 and 1990.
60
Helland & Tabarrok, supra note 43.
61
One of our supplementary tests is a completely independent test using instrumental var-
iables. When jails become overcrowded, judges are pressured to release individuals on their
own recognizance rather than run the risk of setting a bail amount that the defendant might
not be able to secure. (We present evidence in the working paper that bond dealers understand
that overcrowded jails mean less surety business.) We define Ratio as the county jail population
divided by the official jail capacity. A value of Ratio greater than one indicates overcrowding.
We suggest that jail overcrowding is not likely to be correlated with unobservables that affect
FTA and fugitive rates. Using Ratio as an instrumental variable, we again find that surety bail
significantly reduces fugitive rates. For details, see Helland & Tabarrok, supra note 43.
gether, the evidence suggests that we have good estimates that surety release
reduces FTA rates, survival times, and fugitive rates.
VII. Conclusions
When the default was for every criminal defendant to be held until trial,
it was easy to support the institution of surety bail. Surety bail increased the
number of releases relative to the default and thereby spared the innocent
some jail time. Surety release also provided good, albeit not perfect, assurance
that the defendant would later appear to stand trial. When the default is that
every defendant is released, or at least when many people believe that “in-
nocent until proven guilty” establishes that release before trial is the ideal,
support for the surety bail system becomes more complex. How should the
probability of failing to appear and all the costs this implies, including higher
crime rates, be traded off against the injustice of imprisoning the innocent
or even the injustice of imprisoning the not-yet-proven guilty? We cannot
provide an answer to this question, but we can provide a necessary input to
this important debate.
Defendants released on surety bond are 28 percent less likely to fail to
appear than similar defendants released on their own recognizance, and if
they do fail to appear, they are 53 percent less likely to remain at large for
extended periods of time. Deposit bonds perform only marginally better than
release on own recognizance. Requiring defendants to pay their bonds in
cash can reduce the FTA rate similar to that for those released on surety
bond. Given that a defendant skips town, however, the probability of recapture
is much higher for those defendants released on surety bond. As a result,
the probability of being a fugitive is 64 percent lower for those released on
surety bond compared with those released on cash bond. These finding in-
dicate that bond dealers and bail enforcement agents (bounty hunters) are
effective at discouraging flight and at recapturing defendants. Bounty hunters,
not public police, appear to be the true long arms of the law.
bail jumping 119
APPENDIX
TABLE A1
Ordered Probit on Stringency of Release
Variable Coefficient
Local conditions:
Time, in days, to scheduled start of trial ⫺.5821 (.0038)
Local clearance rate (total arrest/total crime) .3957 (.1799)
Defendant is charged with:
Murder .35915** (.051044)
Rape .376661** (.032135)
Robbery .146899** (.028193)
Assault .208538** (.039397)
Other violent crime .048705⫹ (.02932)
Burglary ⫺.10109** (.027554)
Theft ⫺.16676** (.029142)
Other property crime .212824** (.026824)
Drug trafficking ⫺.1147** (.027033)
Other drug crime ⫺.01139 (.041254)
Driving-related crime ⫺.18755** (.016514)
Defendant characteristics:
Age .000854 (.000653)
Female (yesp1) .873055** (.080055)
Active criminal justice status .191588** (.013974)
Previous felonies .244761** (.013558)
Previous failure to appear .123918** (.015137)
Note.—The model includes county and year effects (not shown). Asymptotic standard
errors are in parentheses. There are 58,585 observations.
⫹
Statistically significant at the greater than 10% level.
** Statistically significant at the greater than 1% level (two-sided test).
Bibliography
American Bar Association. Criminal Justice Standards. Washington, D.C.:
American Bar Association, 1985.
Augurzky, Boris, and Schmidt, Christoph M. “The Propensity Score: A
Means to an End.” IZA Discussion Paper No. 271. Bonn: IZA, March
2001.
Ayres, Ian, and Waldfogel, Joel. “A Market Test for Discrimination in Bail
Setting.” Stanford Law Review 46 (1994): 987–1047.
Burton, Bob. Bail Enforcer: The Advanced Bounty Hunter. Boulder, Colo.:
Paladin Press, 1990.
Chamberlin, John A. “Bounty Hunters: Can the Criminal Justice System Live
without Them?” University of Illinois Law Review (1998): 1175–1205.
Clarke, Stevens H.; Freeman, Jean L.; and Koch, Gary G. “Bail Risk: A
Multivariate Analysis.” Journal of Legal Studies 5 (1976): 341–85.
Clines, Francis X. 2001. “Baltimore Gladly Breaks 10-Year Homicide
Streak.” New York Times, January 3, 2001.
Dehejia, Rajeev H., and Wahba, Sadek. “Causal Effects in Non-experimental

Studies: Re-evaluating the Evaluation of Training Programs.” Working
Paper No. 6586. Cambridge, Mass.: National Bureau of Economic Re-
search, June 1998.
Drimmer, Jonathan. “When Man Hunts Man: The Rights and Duties of
Bounty Hunters in the American Criminal Justice System.” Houston Law
Review 33 (1996): 731–93.
Feeney, Floyd. “Foreword.” Pages ix–xiii in Bail Reform in America, edited
by Wayne H. Thomas, Jr. Berkeley: University of California Press, 1976.
Freed, Daniel J., and Wald, Patricia M. Bail in the United States: 1964.
Washington, D.C.: National Conference on Bail and Criminal Justice, 1964.
Goldkamp, John S., and Gottfredson, Michael R. Policy Guidelines for Bail:
An Experiment in Court Reform. Philadelphia: Temple University Press,
1985.
Greene, William H. Econometric Analysis. 4th ed. Upper Saddle River, N.J.:
Prentice Hall, 2000.
Heckman, James J.; Ichimura, Hidehiko; and Todd, Petra E. “Matching as
an Econometric Evaluation Estimator: Evidence from Evaluating a Job
Training Programme.” Review of Economic Studies 64 (1997): 605–54.
Heckman, James J.; Ichimura, Hidehiko; and Todd, Petra. “Matching As an
Econometric Evaluation Estimator.” Review of Economic Studies 65
(1998): 261–94.
Helland, Eric, and Tabarrok, Alexander. “Public versus Private Law Enforce-
ment: Evidence from Bail Jumping.” Working paper. Fairfax, Va.: George
Mason University, 2003. http://mason.gmu.edu/∼atabarro/PublicvsPrivate.pdf.
Howe, Kenneth. “Massachusetts Undertakes Sweeping Reform of Its War-
rants System.” San Francisco Chronicle, June 23, 1999.
Howe, Kenneth, and Hallissy, Erin. “When Justice Goes Unserved:
Thousands Wanted on Outstanding Warrants—but Law Enforcement
Largely Ignores Them.” San Francisco Chronicle, June 22, 1999.
Imbens, Guido W. “The Role of the Propensity Score in Estimating Dose-
Response Functions.” Technical Working Paper No. 237. Cambridge,
Mass.: National Bureau of Economic Research, April 1999.
Kennedy, Spurgeon, and Henry, D. Alan. “Commercial Surety Bail: As-
sessing Its Role in the Pretrial Release and Detention Decision.” Wash-
ington, D.C.: Pretrial Services Resource Center, 1997.
Kimbrough, Carla J. “Crowded Jails Put Squeeze on Bondsmen.” Arkansas
Democrat-Gazette, June 7, 1989.
Landes, William M. “The Bail System: An Economic Approach.” Journal
of Legal Studies 2 (1973): 79–105.
Landes, William M. “Legality and Reality: Some Evidence on Criminal
Procedure.” Journal of Legal Studies 3 (1974): 287–337.
Lechner, Michael. “Identification and Estimation of Causal Effects of Mul-
bail jumping 121
tiple Treatments under the Conditional Independence Assumption.” IZA

Discussion Paper No. 971. Bonn: IZA, December 1999.
Lechner, Michael. “Programme Heterogeneity and Propensity Score Match-
ing: An Application to the Evaluation of Active Labour Market Policies.”
Econometric Society World Congress 2000 Contributed Paper No. 647.
St. Gallen: University of St. Gallen, Swiss Institute for International Eco-
nomics and Applied Economic Research, 2000.
Lecky, George. “Police Name ‘200 Most Wanted.’” Cincinnati Post, Sep-
tember 5, 1997.
Lee, Henry K., and Howe, Kenneth. “Plan to Clear Backlog of Warrants:
Santa Clara County Offering Amnesty to Some.” San Francisco Chronicle,
January 12, 2000.
Leuven, Edwin, and Sianesi, Barbara. “PSMATCH2: Stata Module to Perform
Full Mahalanobis and Propensity Score Matching, Common Support
Graphing, and Covariate Imbalance Testing.” Statistical Software Com-
ponents, No. S432001. Boston: Boston College, Department of Economics,
2003. http://ideas.repec.org/c/boc/bocode/s432001.html.
Myers, Samuel L., Jr. “The Economics of Bail Jumping.” Journal of Legal
Studies 10 (1981): 381–96.
National Association of Pretrial Service Agencies. Performance Standards
and Goals for Pretrial Release.” 2d ed. Wauwatosa, Wis.: National As-
sociation of Pretrial Service Agencies, September 1998.
Prendergast, Jane. “Warrant Amnesty Offered for 1 Day.” Cincinnati En-
quirer, November 19, 1999.
Reynolds, Morgan. “Privatizing Probation and Parole.” Pages 117–28 in
Entrepreneurial Economics: Bright Ideas from the Dismal Science, edited
by Alexander Tabarrok. New York: Oxford University Press, 2002.
Romano, Jay. “Defendants Increasingly Skip Bail.” New York Times, De-
cember 22, 1991.
Rosenbaum, Paul R., and Rubin, Donald B. “The Central Role of the Pro-
pensity Score in Observational Studies for Causal Effects.” Biometrika 70
(1983): 41–55.
Rosenbaum, Paul R., and Rubin, Donald B. “Reducing Bias in Observation
Studies Using Subclassification on the Propensity Score.” Journal of the
American Statistical Association 79 (1984): 516–24.
Rubin, Donald B. “Estimating Causal Effects of Treatments in Randomized
and Nonrandomized Studies.” Journal of Educational Psychology 66
(1974): 688–701.
Rubin, Donald B. “Assignment to Treatment Group on the Basis of a Co-
variate.” Journal of Educational Statistics 2 (1977): 1–26.
Thomas, Wayne H., Jr. Bail Reform in America. Berkeley: University of
California Press, 1976.
Toborg, Mary A. “Bail Bondsmen and Criminal Courts.” Justice System
Journal 8 (1983): 141–56.
U.S. Department of Justice, Bureau of Justice Statistics. “Felony Defendants

in Large Urban Counties, 1996.” Washington, D.C.: Bureau of Justice
Statistics, 1999. http://www.ojp.usdoj.gov/bjs/pub/pdf/fdluc96.pdf.

Tabarrok Study

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Tabarrok Study

Uploaded by

Copyright:

Available Formats

THE FUGITIVE: EVIDENCE ON PUBLIC VERSUS

PRIVATE LAW ENFORCEMENT FROM

A pproximately one-quarter of all released felony defendants fail to appear

[Journal of Law and Economics, vol. XLVII (April 2004)]

Defendants who fail to appear impose significant costs on others. Direct

a defendant fails to appear at least once, the probability that a defendant

II. History of Pretrial Release and Incentive Effects

significant efforts were made, beginning in the 1960s, to develop alternatives

union membership, previous arrests, and so on.18 In addition, bond dealers

III. The Matching Model with Multiple Treatments

formed into a single-dimension problem of matching on Pr(T p 1FX), where

v0m,l p E(Y m ⫺ Y lFT p m) p E(Y mFT p m) ⫺ E(Y lFT p m), (1)

where v0m,l denotes the effect of treatment m rather than l.

If this assumption is valid, we can use the conditional propensity score to

In practice, the conditional propensity score, p mFml (x), is computed indirectly

IV. Data and Descriptive Statistics

Own Deposit Cash Surety Emergency

A. Propensity Scores from Ordered Probit

decisions.42 Other, nonindividual variables include the police clearance rate,

Figure 1B gives the box-and-whiskers plot for the pairwise (conditional)

Own Deposit Cash Surety

C. Estimated Treatment Effects: Failure to Appear

D. Estimated Treatment Effects: The Fugitive

Own Deposit Cash Surety

to purposively abscond. If this is correct, we ought to see a positive correlation

E. Kaplan-Meier Estimation of Failure-to-Appear Duration

F. Comparison with Counties in States That Have

VI. Looking for Unobservable Variables

egies; a number of alternative strategies, described briefly below, are devel-

Surety versus Surety versus Surety versus

Note.—Own-own indicates first release on own recognizance

Dehejia, Rajeev H., and Wahba, Sadek. “Causal Effects in Non-experimental

tiple Treatments under the Conditional Independence Assumption.” IZA

U.S. Department of Justice, Bureau of Justice Statistics. “Felony Defendants

You might also like